The Köppen climate classification scheme was first introduced by Wladimir Köppen (Köppen, 1936). It is one of the earliest quantitative classification systems of Earth's climates. Its modification, the Köppen–Geiger classification (KGC), was first published in 1936 (Köppen, 1936), developed by Wladimir Köppen and Rudolf Geiger. KGC identifies climates based on their effects on plant growth from the aspects of warmth and aridity and classifies climate into five main climate classes and 30 subtypes (Rubel and Kottek, 2011). The five main climate zones distinguish between plants of the tropical climate zone (A), the arid climate zone (B), the temperate climate zone (C), the boreal climate zone (D), and the polar climate zone (E) (Sanderson, 1999). All main climate zones are thermal zones except the arid (B) climate zone, which is defined based on precipitation threshold.

This research followed the Köppen–Geiger climate classification as described in Kottek et al. (2006) and Rubel and Kottek (2010). This latest version of the KGC scheme was first presented by Geiger (1961) (Table 2). Several existing Köppen–Geiger climate map products, including Peel et al. (2007), Kriticos et al. (2012), and Beck et al. (2018), applied the KGC scheme modified following Russell (1931). Russell (1931) adjusted the definition of the boundary of temperate (C) and boreal (D) climate zones using the coldest monthly temperature > 0 ∘C instead of > -3 ∘C. This threshold was proposed because the 0 ∘C line fits the distribution of the topographical features and vegetation in the western United States, where at that time meteorological stations were sparsely distributed (Jones, 1932). However, the application of 0 ∘C boundary to the global climates has not been validated. Therefore, this research did not utilize the Russell (1931) modification and followed the latest version KGC proposed by Geiger (1961).

Due to limited number of available observational datasets with high resolution and long-term continuous temporal coverage, the research implemented the delta method by applying a delta change or change factor (Hay et al., 2000; Wilby and Wigley, 1997) onto the WorldClim historical observations (Fick and Hijmans, 2017) to achieve 30-year average climatology data with a 1 km resolution based on the CRU, UDEL, and GPCC datasets. The delta method is a statistical downscaling method that assumes that the relationship between climatic variables remains relatively constant at local scale (Wilby and Wigley, 1997). We applied the delta method to downscale the long-term (30-year) mean climates using coarse-resolution monthly climatology datasets. The delta changes or change factors are calculated as the differences between the 30-year long-term means of temperature or precipitation of baseline (1970–2000) and present-day climates. The delta method comprises the following four steps: (1) calculate 30-year averages for the baseline (1970–2000) and present day of monthly temperature and precipitation, (2) calculate anomaly for precipitation and delta for temperature, (3) apply thin-plate spline (TPS) interpolation to create 1 km surface of precipitation anomaly and temperature delta, and (4) multiply anomaly or add delta to historical climates based on WorldClim dataset (Fig. 1).

First, using monthly time series from the CRU, UDEL, and GPCC datasets, we calculated 30-year means as a baseline (1970–2000) for each climatology dataset and each variable. We used 1970–2000 as the baseline period, for consistency with WorldClim Historical Climate Data V2. Next, we calculated 30-year means for each month and each 30-year present-day period in 1979–2013. We then calculated anomalies as proportional differences between present day and baseline in total precipitation and delta as the difference in temperature. To derive 30 arcsec (1 km) anomaly or delta surfaces, we applied thin-plate spline (TPS) interpolation (Franke, 1982; Schempp et al., 1977; Craven and Wahba, 1978) to precipitation anomaly and temperature delta. TPS has been widely used in climate science (Hijmans et al., 2005; Navarro-Racines et al., 2020) as it produced a smooth and continuous surface, which is infinitely differentiable. Last, we multiplied the change factor or added the delta to the WorldClim (1970–2000) data to get downscaled present-day monthly climate data.

Our future Köppen–Geiger map series are based on an ensemble of maps derived from the CCAFS bias-corrected and downscaled climate projections, which include 35 CMIP5 GCMs and 4 RCPs (Navarro-Racines et al., 2020). Large misclassifications exist within the GCMs as detected in previous assessments of large areas ranging between 20 %–50 % of the total land area (Cui et al., 2021a). Deficiencies in model physics are also more likely to contribute to uncertainties in the maps than grid size or reference dataset limitations (Tapiador et al., 2019). Multi-model mean and delta-change methods can mitigate the bias effects from the threshold-based classification scheme and have been utilized to simulate better results of climate classification (Hanf et al., 2012). Therefore, we chose the CCAFS bias-corrected and downscaled CMIP5 projections (Navarro-Racines et al., 2020) to reduce the amplified errors due to uncertainty of climate projections. Navarro-Racines et al. (2020) interpolated anomalies of original GCM outputs using thin plate spline spatial interpolation to achieve a baseline climate with a 1 km surface. Then they applied the delta method to the interpolated baseline climates to correct the model biases (Hay et al., 2000; Ho et al., 2012).

The historical Köppen–Geiger climate classification map series was generated using the highest confidence class from an ensemble of maps using all combinations of surface air temperature and precipitation products (Fig. 2), as described in Beck et al. (2018). The highest confidence was given to the most common climate class for each grid cell. The final historical climate map series were derived using the climate class with the highest level of confidence in an ensemble of 3 × 3 = 9 classification maps based on combinations of the three precipitation datasets (CRU, UDEL, and CHELSA) and three surface air temperature datasets (GPCC, UDEL, and CHELSA). To further test the sensitivity of the method using the climate with the highest level of agreement, we incorporated another data integration method using the mean of multiple datasets. We quantified the degree of confidence placed in the Köppen–Geiger climate map series using the degree of confidence at the grid cell level calculated by dividing the occurrence frequency of the climate class with the highest level of agreement by the ensemble size. The calculated confidence level can be viewed as the agreement degree in classification derived from different climatology datasets.

The future Köppen–Geiger climate classification map series under four RCPs were derived based on the most common climate class from an ensemble of future climate maps. We generated a future Köppen–Geiger climate classification map for each climate model projection, using the CCAFS bias-corrected and downscaled CMIP5 GCM dataset. For example, the future Köppen–Geiger climate classification map series under RCP8.5 was derived from an ensemble of 30 maps based on 30 CMIP5 models. The level of confidence was estimated using the ratio between the frequency of the climate class with the highest level of agreement in the future map results and the ensemble size.

We validated the historical climate maps using the station observations from Global Historical Climatology Network-Daily (GHCN-D) (Menne et al., 2012) and Global Summary of the Day (GSOD) database (National Climatic Data Center et al., 2015) as reference data. The GHCN-D dataset provides daily climate data over global land areas and contains records from over 80 000 weather stations worldwide, about one-third of which have both temperature and precipitation data available (Menne et al., 2012). The GSOD dataset includes global daily summary data over 9000 stations, of which the historical data from 1973 are the most complete (National Climatic Data Center et al., 2015). For each station, time series of monthly temperature and precipitation were calculated from the daily observations with months with <15 daily values discarded. Then if ≥ 6 months are present, monthly climatologies were subsequently generated by averaging the monthly means for the given 30-year period. We removed duplicate stations in the two datasets and discarded stations with gap years or missing data in the given 30 years. For each station and each 30-year period, we applied the Köppen–Geiger climate classification, and then we evaluated overall classification performance for each climate map using total accuracy, which is defined as the percentage of correct classes and average precision, which is the averaged fraction of correct classification for all climate classes.

Using the same validation datasets and station selection process, we also evaluated the previous climate maps from Beck et al. (2018), Kriticos et al. (2012), Peel et al. (2007), and Kottek et al. (2006). We applied the same Köppen–Geiger climate classification criteria described in the previous studies to assess the overall accuracy of the map products. To further validate the climate classification results, we performed sensitivity analysis on the data integration method, the climate classification timescale, and climatology dataset input, using the same validation datasets from GHCN-D and GSOD. In addition, we compared the climate classification results with forest cover and elevation maps and with the two high-resolution comparable climate map products, Beck et al. (2018) (1 km) and Kriticos et al. (2012) (0.167∘), at regional and continental scales. The forest cover map we used is the 2000 30 m Landsat-based forest cover map (Hansen et al., 2013). The elevation data are from the NASA SRTM Digital Elevation 30 m data (Farr et al., 2007).

Global map series of the Köppen–Geiger climate classification for historical periods and associated corresponding confidence levels are shown in Fig. 3. Based on the distribution of confidence level, over 90 % of the land area exhibits a high level of confidence as classification results based on different climate data show excellent agreement. Relatively lower confidence level and large discrepancy in classification results are found in particular in mountainous regions such as the Andes Mountains, Rocky Mountains, Tibetan Plateau, and major climate transitional zones located in the midlatitude and high latitudes of Northern Hemisphere, central Africa, and central Asia.

Regional distributions of climatic conditions are largely created by local variation in topography in rugged terrain (Dobrowski et al., 2013; Franklin et al., 2013). The climate classification and confidence level maps of mountainous areas of the central Rocky Mountains and Tibetan Plateau are shown in Figs. 4 and 5, respectively. For each combination of precipitation and surface air temperature datasets, we generated a Köppen–Geiger climate classification map (see Figs. 4a and 5a for 1979–2008 maps for the central Rocky Mountains and Tibetan Plateau). The final Köppen–Geiger classification map is derived based on the most common climate type among all the climate maps (Figs. 4b and 5b). We then calculated corresponding confidence levels to quantify the uncertainty in the classification maps (Figs. 4c and 5c). The uncertainty in climate classification in mountainous areas is attributed to the uncertainty existing in climate data, especially precipitation data. In rugged terrain, CHELSA precipitation data show more detailed precipitation patterns, causing disagreement in classification results of the third-level climate classes which depict precipitation seasonality.

We validated the historical climate maps using the station observations from Global Historical Climatology Network-Daily (GHCN-D) (Menne et al., 2012) and Global Summary of the Day (GSOD) database (National Climatic Data Center et al., 2015). Figure 6 shows the small-scale distributions of total accuracy and average precision for historical Köppen–Geiger climate map series with 10∘ grid cells. Due to uneven distributions of weather stations, remote areas in the Pacific islands, central Africa, and Amazon forest suffer from a lack of station observations or underrepresented validation results.

We summarized the overall accuracy, average precision, and confidence levels for each continent and the whole globe (Table 3). The global overall classification accuracy of the historical Köppen–Geiger climate maps is estimated to be 82.39 %, with the lowest in South America (68.58 %) and highest in Oceania (92.01 %). The global average precision, which is calculated as averaged fraction of correct classification for all climate classes, is 73.33 %. Similar to overall accuracy, South America has the lowest precision level, equal to 66.35 %, and Oceania the highest, 92.23 %. Having a good correspondence with accuracy and precision values, the continental average confidence levels range from 91.55 % to 94.93 %, and the global level is 92.90 % (Table S2). Overall, the spatial patterns of total accuracy and average precision show good correspondence with classification confidence levels (Fig. 3), indicating a potential of confidence level to represent classification uncertainty.

Using the same validation datasets from GHCN-D and GSOD, we tested sensitivity of the climate map series using different combinations of temperature and precipitation datasets and different methods of data integration (Table 4). Results indicated an average total accuracy of the 1 km Köppen–Geiger classification maps generated with all the CHELSA and downscaled CRU, GPCC, and UDEL datasets and with only downscaled CRU, GPCC, and UDEL datasets as 82.39 % and 81.17 %, respectively. Using the mean of multiple datasets, which can potentially reduce the data bias, led to better classification results. We estimated the total accuracy of the previous high-resolution Köppen–Geiger climate map products using the same validation datasets. We applied the same classification system described in the previous studies and the same time period of the previous climate map product to process the station observation data and estimate their overall accuracy. Compared with the previous high-resolution Köppen–Geiger climate map products, Beck et al. (2018) and Kriticos et al. (2012), the newly generated Köppen–Geiger climate map series showed greater accuracy in total.

We conducted sensitivity analysis of the Köppen classification scheme and tested multiple timescales, 10 years, 20 years, and 30 years. The selection criteria of station observations were adjusted accordingly based on the timescale utilized. Accuracy results exhibited decreasing accuracy for shorter timescales (Fig. 7). Further, we estimated the total accuracy for the Köppen–Geiger climate classification maps from previous studies, Beck et al. (2018) Kriticos et al. (2012), Peel et al. (2007), and Kottek et al. (2006), using the same validation dataset and consistent Köppen–Geiger climate classification scheme the corresponding study applied. The validation results demonstrate that the new Köppen–Geiger maps have comparatively higher overall accuracy than all the previous studies.

At the regional and continental scale, we compared our Köppen–Geiger climate classification maps with previous map products for regions with large spatial gradients in climates, including central and eastern Africa, Europe, and North America, and regions with sharp elevation gradients, including the Tibetan Plateau, central Rocky Mountains, and central Andes (Fig. 8). We compared the new 1 km Köppen–Geiger climate classification maps from our study for time periods of 1980–2009 and 1984–2013 with the high-resolution Köppen–Geiger maps from two previous studies, Beck et al. (2018), which has a resolution of 1 km and temporal coverage of 1980–2016, and Kriticos et al. (2012), which has a resolution of 0.0167∘ and covers 1961–1990. The Köppen classifications demonstrate good correlation with natural landscape distributions (Belda et al., 2014; Köppen, 1936; Trewartha, 1954). To show the agreement between the improved Köppen–Geiger climate classification maps and regional landscape distributions, we also showed maps of forest cover and elevation distribution for these regions. Figure 8 illustrates the enhanced regional details of the maps.

Compared with the Köppen–Geiger climate maps from previous studies with only one time period, the series of the Köppen–Geiger climate maps from our study demonstrate the ability to capture recent changes in spatial distributions of climate zones. For example, our maps can detect the significant changes in the climate zones specifically driven by the accelerated global warming since the 1980s, for example, the poleward movements of boreal (D) and polar (E) climates in high latitudes in North America shown in the comparison between the 1980–2009 and 1984–2013 Köppen–Geiger climate maps (Fig. 8d). Another example is the expansion of savanna (Aw) climate into the temperate (Cw) climate zone, witnessed in central Africa (Fig. 8e).

Another improvement of the new series of the Köppen–Geiger climate maps is the application of a threshold of -3 ∘C as the boundary of temperate (C) and boreal (D) climate zones, which show better agreement with global boreal forest distributions at the regional scale compared with the Russell (1931) modification of 0 ∘C, which Beck et al. (2018) and Kriticos et al. (2012) utilized. Based on the comparison results of the Köppen climate zones and the biome classifications from the World Wildlife Federation (Rohli et al., 2015b), the boreal (D) climate zone largely corresponds to the distribution of boreal forest (Cui et al., 2021a). For example, evidenced in Fig. 8c, the new Köppen–Geiger climate classification maps from our study show better agreement with the boreal forest in the Carpathian Mountains across central and eastern Europe than Beck et al. (2018) and Kriticos et al. (2012). Figure 8d also shows good agreement of the northern boundary of the boreal (D) climate zone in the northern part of Quebec in Canada with the boundary of Canada's boreal forest.

Moreover, the new Köppen–Geiger maps can show an accurate depiction of important topographic features over the regions with complex topography. For example, the topo-climate of the Himalayas' southern front determined by the mountain ranges is represented with more detail in the new Köppen–Geiger maps compared with Beck et al. (2018) and Kriticos et al. (2012) (Fig. 8b). The abrupt changes in climate along the edges of the Andes mountains are also well described in the new maps (Fig. 8f).

In addition, the distribution of tropical (A), temperate (C), and boreal (D) climate zones in the new Köppen–Geiger maps correspond closely with tree lines in the forest cover maps. The temperate (C) and boreal (D) climate distributions based on the Köppen–Geiger maps show a better agreement with the forest distributions of the middle and southern Rocky Mountains than Beck et al. (2018) and Kriticos et al. (2012) (Fig. 8a). For another example, the boundaries of the tropical rainforest in central Africa and South America are clearly delineated in the new Köppen–Geiger maps (Fig. 8e and f).

Beyond the Köppen–Geiger climate classification maps, we calculated a set of bioclimatic variables from the monthly climate data (see full list in Table 5). The bioclimatic variables at 1 km spatial resolution can capture regional environmental variations in particular in mountainous areas and areas with strong climate variations. These bioclimatic variables can be used in studies of environmental, agricultural, and biological sciences, for example, development of species distribution modeling and assessment of biological impacts induced by climate change. The variables provide descriptions of annual averages and seasonality of climates. The warmest half year or the coldest half year is defined as the period of the warmest 6 months or the coldest 6 months.

We validated the bioclimatic variables from different datasets with station data from GHCN-D (Menne et al., 2012) and the GSOD database (National Climatic Data Center et al., 2015) (Fig. 9). We calculated a linear regression model for the 12 bioclimatic variables for each 10∘ grid cell (Fig. 10). The 30-year average mean annual temperature (MAT) from the CHELSA dataset shows the overall highest fit with station data, with CRU, and UDEL datasets show a smaller but still strong correlation with station data. The 30-year average mean annual precipitation (MAP) estimates from the GPCC, UDEL, and CHELSA datasets have considerable uncertainties, indicated by relatively low correlation with station observations. In current precipitation datasets, there is a varied degree of discrepancy in annual estimates over multiple timescales (Sun et al., 2018).

Future Köppen–Geiger climate classification maps under RCP8.5 and associated confidence levels are shown in Fig. 11. Indicated by confidence levels, there exist larger uncertainties in the final future climate maps than historical maps, particularly at midlatitudes and high latitudes. The climate map for the future period of 2070–2099 shows the largest uncertainty compared with the other future periods.

Future climate classifications derived from the diverse GCM projections for four RCPs, which are inherently uncertain (Winsberg, 2012; Gleckler et al., 2008), provide a proxy of global distributions of climatic conditions and can represent potential spatial changes in climate zones under global warming. The large uncertainty and strong disagreement in projected climate classification maps at high latitudes and in regions with rugged terrain can be indicated by relatively low confidence levels. Figures 12 and 13 show the future Köppen–Geiger climate classification maps based on GCM projections under RCP8.5 and associated confidence levels for the central Rocky Mountains and Tibetan Plateau. We generated a future Köppen–Geiger climate classification map for each bias-corrected and downscaled CMIP5 GCM projection (see Figs. 12a and 13a for 2070–2099 maps for the central Rocky Mountains and Tibetan Plateau). Noticeable regional changes in climate zones have been projected by comparing the 2070–2099 and 1979–2008 climate classification maps (see Fig. 12b and c for the central Rocky Mountains and Fig. 13b and c for the Tibetan Plateau).

Changes in climatic conditions under global warming have significant impacts on biodiversity and ecological systems. Area changes of climate zones can indicate spatial shrinkage or expansion of analogous climatic conditions, potentially implying threats for species range contraction or opportunities for range expansion (Cui et al., 2021a). To examine the area changes of climate zones, we calculated the total area covered by each climate type for each historical and future period under the high-emission RCP8.5 scenario (Fig. 14). Our results of changes in area occupied by different climate zones demonstrate good agreement with results from previous studies (Chan and Wu, 2015). Results show that accelerated anthropogenic global warming since the 1980s has caused large-scale changes in climate zones, and shifts into warmer and drier climates are projected in this century. The tropical and arid climates are expanding into large areas in midlatitudes, whereas the high-latitude climates will experience significant area shrinkage.